supervision/docs/notebooks/annotate-video-with-detections.ipynb

Annotate Video with Detections

---

One of the most common requirements of computer vision applications is detecting objects in images and displaying bounding boxes around those objects. In this cookbook we'll walk through the steps on how to utilize the open source Roboflow ecosystem to accomplish this task on a video. Let's dive in!

Before you start

Let's make sure that we have access to GPU. We can use nvidia-smi command to do that. In case of any problems navigate to Edit -> Notebook settings -> Hardware accelerator, set it to GPU, and then click Save.

In [1]:

!nvidia-smi

Fri Feb 23 03:15:00 2024       
+---------------------------------------------------------------------------------------+
| NVIDIA-SMI 535.104.05             Driver Version: 535.104.05   CUDA Version: 12.2     |
|-----------------------------------------+----------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |         Memory-Usage | GPU-Util  Compute M. |
|                                         |                      |               MIG M. |
|=========================================+======================+======================|
|   0  Tesla V100-SXM2-16GB           Off | 00000000:00:04.0 Off |                    0 |
| N/A   33C    P0              24W / 300W |      0MiB / 16384MiB |      0%      Default |
|                                         |                      |                  N/A |
+-----------------------------------------+----------------------+----------------------+
                                                                                         
+---------------------------------------------------------------------------------------+
| Processes:                                                                            |
|  GPU   GI   CI        PID   Type   Process name                            GPU Memory |
|        ID   ID                                                             Usage      |
|=======================================================================================|
|  No running processes found                                                           |
+---------------------------------------------------------------------------------------+

Installing Dependencies

In this cookbook we'll be utilizing the open source packages Inference and Supervision to accomplish our goals. Let's get those installed in our notebook with pip.

In [ ]:

!pip install -q inference-gpu "supervision"

Download a Video Asset

First, let's download a video that we can detect objects in. Supervision comes with a great utility called Assets to help us hit the ground running. When we run this script, the video is saved in our local directory and can be accessed with the variable path_to_video.

In [ ]:

from supervision.assets import download_assets, VideoAssets

# Download a supervision video asset
path_to_video = download_assets(VideoAssets.PEOPLE_WALKING)

As a result, we've downloaded a video. Let's take a look at the video below. Keep in mind that the video preview below works only in the web version of the cookbooks and not in Google Colab.

Detecting Objects

For this example, the objects in the video that we'd like to detect are people. In order to display bounding boxes around the people in the video, we first need a way to detect them. We'll be using the open source Inference package for this task. Inference allows us to quickly use thousands of models, including fine tuned models from Roboflow Universe, with a few lines of code. We'll also utilize a few utilities for working with our video data from the Supervision package.

In [ ]:

import supervision as sv
from inference.models.utils import get_roboflow_model


# Load a yolov8 model from roboflow.
model = get_roboflow_model("yolov8s-640")

# Create a frame generator and video info object from supervision utilities.
frame_generator = sv.get_video_frames_generator(path_to_video)

# Yield a single frame from the generator.
frame = next(frame_generator)

# Run inference on our frame
result = model.infer(frame)[0]

# Parse result into detections data model.
detections = sv.Detections.from_inference(result)

# Pretty Print the resulting detections.
from pprint import pprint
pprint(detections)

Detections(xyxy=array([[1140.,  951., 1245., 1079.],
       [ 666.,  648.,  745.,  854.],
       [  34.,  794.,  142.,  990.],
       [1140.,  505., 1211.,  657.],
       [ 260.,  438.,  332.,  612.],
       [1413.,  702., 1523.,  887.],
       [1462.,  472., 1543.,  643.],
       [1446.,  318., 1516.,  483.],
       [ 753.,  451.,  821.,  623.],
       [ 924.,  172.,  983.,  307.],
       [1791.,  144., 1852.,  275.],
       [  93.,  132.,  146.,  251.],
       [ 708.,  240.,  765.,  388.],
       [ 200.,   44.,  267.,  161.],
       [1204.,  131., 1255.,  266.],
       [ 569.,  267.,  628.,  408.],
       [1163.,  150., 1210.,  280.],
       [ 799.,   78.,  847.,  204.],
       [1690.,  152., 1751.,  283.],
       [ 344.,  495.,  396.,  641.],
       [1722.,   77., 1782.,  178.]]),
           mask=None,
           confidence=array([0.83215541, 0.80572134, 0.7919845 , 0.7912274 , 0.77121079,
       0.7599591 , 0.75711554, 0.75494027, 0.73076195, 0.71452248,
       0.69572842, 0.65269446, 0.63952065, 0.62914598, 0.61361706,
       0.5968492 , 0.55311316, 0.5470854 , 0.54070991, 0.52209878,
       0.41217673]),
           class_id=array([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]),
           tracker_id=None,
           data={'class_name': array(['person', 'person', 'person', 'person', 'person', 'person',
       'person', 'person', 'person', 'person', 'person', 'person',
       'person', 'person', 'person', 'person', 'person', 'person',
       'person', 'person', 'person'], dtype='<U6')})

First, we load our model using the method get_roboflow_model(). Notice how we pass in a model_id? We're using an alias here. This is where we can pass in other models from Roboflow Universe like this rock, paper, scissors model utilizing our roboflow api key.

model = get_roboflow_mode(
    model_id="rock-paper-scissors-sxsw/11",
    api_key="roboflow_private_api_key"
)

If you don't have an api key, you can create an free Roboflow account. This model wouldn't be much help with detecting people, but it's a nice exercise to see how our code becomes model agnostic!

We then create a frame_generator object and yeild a single frame for inference using next(). We pass our frame to model.infer() to run inference. After, we pass that data into a little helpfer function called sv.Detections.from_inference() to parse it. Lastly we print our detections to show we are in fact detecting a few people in the frame!

Annotaing the Frame with Bounding Boxes

Now that we're detecting images, let's get to the fun part. Let's annotate the frame and display the bounding boxes on the frame.

In [5]:

# Create a bounding box annotator object.
box_annotator = sv.BoxAnnotator()

# Annotate our frame with detections.
annotated_frame = box_annotator.annotate(scene=frame.copy(), detections=detections)

# Display the frame.
sv.plot_image(annotated_frame)

Notice that we create a box_annotator variable by initalizing a BoxAnnotator. We can change the color and thickness, but for simplicity we keep the defaults. There are a ton of easy to use annotators available in the Supervision package other than a bounding box that are fun to play with.

Saving Bounding Boxes to a Video

Let's wrap up our code by utilizing a VideoSink to draw bounding boxes and save the resulting video. Take a peak at the final code example below. This can take a couple minutes deppending on your runtime and since since we're processing a full video. Feel free to skip ahead to see the resulting video.

In [ ]:

from tqdm import tqdm

# Create a video_info object for use in the VideoSink.
video_info = sv.VideoInfo.from_video_path(video_path=path_to_video)

# Create a VideoSink context manager to save our frames.
with sv.VideoSink(target_path="output.mp4", video_info=video_info) as sink:

    # Iterate through frames yielded from the frame_generator.
    for frame in tqdm(frame_generator, total=video_info.total_frames):

        # Run inference on our frame.
        result = model.infer(frame)[0]

        # Parse the result into the detections data model.
        detections = sv.Detections.from_inference(result)

        # Apply bounding box to detections on a copy of the frame.
        annotated_frame = box_annotator.annotate(
            scene=frame.copy(),
            detections=detections
        )

        # Write the annotated frame to the video sink.
        sink.write_frame(frame=annotated_frame)

In the code above we've created avideo_info variable to pass information about the video to our VideoSink. The VideoSink is a cool little context manager that allows us to write_frames() to a video output file. We're also optionally using tqdm to display a progress bar with a % complete. We only scratched the surface of all of the customizable Annotators and additional features that Supervision and Inference have to offer. Stay tuned for more cookbooks on how to take advantage of them in your computer vision applications. Happy building! 🚀